SSC Annual Meeting, June 2007

Proceedings of the Survey Methods Section

A WEIGHT SMOOTHING METHOD FOR DEALING WITH

STRATUM JUMPERS IN BUSINESS SURVEYS
Jean-Franois Beaumont and Louis-Paul Rivest1
ABSTRACT
The problem of stratum jumpers in stratified business surveys is mainly due to inaccurate size information at the time of
sampling, which may result in assigning a large design weight to a large size unit. Such units may be quite influential and
substantially inflate the variance of the estimates. They can be viewed as being outliers with respect to the implicit model used at
the sampling stage. Therefore, standard robust estimation techniques can be used to handle this problem. Instead, we will
consider an alternative weight smoothing method and illustrate its application using data from the Canadian Workplace and
Employee Survey.
KEY WORDS: Extreme weight; Model; Robust estimation; Smoothed weight.

RSUM
Le problme des units "sauteuses" de strate dans les enqutes stratifies auprs des entreprises est principalement d des
informations imparfaites sur la taille de l'entreprise au moment de l'chantillonnage, ce qui peut conduire assigner un grand
poids de sondage une unit de grande taille. De telles units peuvent avoir une grande influence et augmenter significativement
la variance des estimations. Elles peuvent tre vues comme tant des donnes aberrantes par rapport au modle implicite utilis
l'tape d'chantillonnage. Par consquent, les techniques usuelles d'estimation robuste peuvent tre utilises pour tenir compte de
ce problme. Au lieu de ces techniques, nous considrerons une mthode alternative de lissage des poids et illustrerons son
application au moyen des donnes de l'Enqute canadienne sur le milieu du travail et les employs.
MOTS CLS : Poids extrme; Modle; Estimation robuste; Poids liss.

1. INTRODUCTION
In many business surveys, the population of businesses is stratified by region, industry type and size group. The latter is
defined using some measure of business size available on the sampling frame, such as the number of employees or the
revenue of the business. Then, the sample is usually selected by stratified simple random sampling without replacement.
For efficiency considerations, a large selection probability i (and thus a small design weight, d i = 1 i ) is usually
assigned to a unit i of large size on the sampling frame while a small selection probability (and thus a large design weight)
is assigned to a unit of small size. This strategy is justified on the grounds that business survey variables are usually
highly skewed and that there is usually a positive correlation between the size measure and the main variables of interest
so that large values of these variables are expected to be assigned to a small design weight.
At the time of collection, we often observe discrepancies between the information available at the design stage and the
same information collected from the respondent. These discrepancies can be explained by errors on the sampling frame,
which are partly due to outdated information, and the time lag between sampling and collection. They become
problematic when a unit that is thought to be of small size at the design stage is actually found to be a large unit. Such
units are sometimes called stratum jumpers because they would have been assigned to another stratum had the correct
1

Jean-Franois Beaumont, Statistical Research and Innovation Division, Statistics Canada, 16th floor, R.H. Coats Building, Ottawa, Canada, K1A
0T6 (Jean-Francois.Beaumont@statcan.ca) and Louis-Paul Rivest, Departement de mathmatiques et de statistique, Universit Laval, Cit
universitaire, Qubec (Qubec), Canada G1K 7P4 (lpr@mat.ulaval.ca).

information been available at the time of design. A consequence of this problem is that some units with large values of the
variables of interest are unfortunately assigned large design weights, which may result in inefficient design-based
estimators. It may actually occur that a few stratum jumpers account for an important percentage, say 20% to 30%, of the
estimate of a population total.
At the design stage, the potential impact of stratum jumpers can be reduced to some extent by controlling the maximum
design weight to be smaller than a certain threshold. This will usually imply departing from optimal stratification and/or
allocation. Rivest (1999) proposed a method for dealing with stratum jumpers, which worked well empirically. It reduces
the maximum design weight by a large factor. No matter how carefully the sampling design is chosen, it is likely that the
problem will not be completely eliminated as the stratum jumpers occur in a haphazard way.
As an example, suppose that there are two design strata A and B. Stratum A has 9 selected units considered to be of large
size at the design stage, which are assigned a design weight of 1, while stratum B has 41 selected units considered to be of
small size at the design stage, which are assigned a design weight of 31. At the collection stage, we observe that one of the
41 units with a large weight is actually a large size unit so that the collection stratum is different from the design stratum
for this unit, which is thus called a stratum jumper. Table 1 summarizes the above information. The stratum jumper is
found in the middle row of this table.
Table 1. Example showing a stratum jumper
Collection stratum Design stratum Number of units Design weight
A
A
9
1
A
B
1
31
B
B
40
31
50
1280
Sum over the sample units
Assume that the collection strata are homogeneous with respect to the variable of interest y (or at least more homogeneous
than the design strata) so that collection stratum A contains large size units associated with large y-values while collection
stratum B contains small size units associated with small y-values. On the one hand, the stratum jumper can be viewed as
a unit with a large y-value compared to the other 40 units in the same design stratum, which all have the same design
weight. Therefore, standard outlier-robust techniques, such as winsorization or M-estimation, could be used to handle this
stratum jumper. On the other hand, the stratum jumper can also be viewed as a unit with a large design weight compared
to the other 9 units in the same collection stratum although it may have a similar y-value. This is the view we take in
Section 3 with our weight smoothing approach.
Before describing our weight smoothing approach, let us first briefly discuss an alternative approach to handling the
potential problem of extreme design weights; namely, winsorizing the largest design weights (e.g., Potter, 1990; and Liu,
Ferraro, Wilson and Brick, 2004). This actually seems to be the most popular method whenever something is done to deal
with the problem of large design weights. Using this approach in the example given in Table 1 would lead to reducing the
design weight of the stratum jumper but also of all other units in the same design stratum, which may be less appealing
and may reduce efficiency. The main challenge with this method is to determine an appropriate winsorization cut-off.
Several methods, sometimes more or less ad hoc, have been considered to address this issue, including the use of outlier
detection techniques. One data-driven method is to choose the cut-off so as to minimize an estimate of the design Mean
Squared Error (MSE). Unfortunately, this leads to a different cut-off for each variable of interest y, and thus a different
winsorized weight for each y-variable. This is not convenient in multipurpose surveys and, thus, a compromise
winsorization cut-off is needed. Also, it is worth noting that extreme design weights may not cause any problem if there is
no stratum jumper and the design strata are homogeneous.
In Section 2, we briefly describe a weight smoothing approach, proposed by Beaumont (2008), to deal with the general
problem of variable design weights, including the presence of extreme weights. This approach is adapted to handle the
problem of stratum jumpers in stratified business surveys in Section 3. Unlike winsorization of design weights, a
compromise smoothed weight is obtained naturally when there is more than one y-variable. In Section 4, the potential
benefits of weight smoothing are illustrated using the data of the Canadian Workplace and Employee Survey (CWES).
Finally, in Section 5, we conclude with a brief summary and discussion.

2. A GENERAL WEIGHT SMOOTHING APPROACH

We consider the problem of estimating the population total Ty =

iU

yi for a population U of size N. A sample s is

selected from the population according to a probability sampling design p ( s Z) , where Z = ( z 1 ,..., z N ) and z i is a
vector of design variables (e.g., the size measure, region or industry) for the ith population unit. Let us also use the
notation I = ( I 1 ,..., I N ) , X = (x1 ,..., x N ) and Y = ( y1 ,..., y N ) , where xi is a potential vector of auxiliary variables for
the ith population unit that may be used for calibration purposes and I i is the sample inclusion indicator of unit i; i.e.

I i = 1 if the ith population unit is selected in the sample s and I i = 0 , otherwise. It should be kept in mind that the design
weight d i = 1 i is a function of Z only and should thus be written d i (Z) . Nevertheless, we still denote it by d i for
convenience.
The basic idea underlying weight smoothing consists of first viewing the design weights as random and then using the
model :

E ( d i | I , X, Y ) = g s ( x i , y i ; s ) ,

(1)

for i s , where g s (.;.) is some function that may be sample-dependent and s is a vector of unknown model
parameters to be estimated from sample data. Specific models for the design weights are given in Pfeffermann and
Sverchkov (1999) and Beaumont (2008). Note that nothing precludes y from being a vector so that the problem of finding
a smoothed weight in multipurpose surveys boils down to considering more explanatory variables in model . If

~
~
d i = g s (x i , y i ; s ) were known, we would obtain the smoothed estimator T ySDB by replacing the design weights d i by
~
the smoothed weight d i in a design-based estimator TyDB , such as the Horvitz-Thompson estimator or a calibration
estimator (Deville and Srndal, 1992). For instance, if TyDB is the Horvitz-Thompson estimator, i.e., TyDB = is d i yi ,
~
~
then T ySDB = is d i y i . The role of model is to remove the unnecessary variability in the design weights. Beaumont
~
(2008) showed that T ySDB is unbiased and never less efficient than the corresponding design-based estimator TyDB under
the model and the sampling design.

Since d i = g s (x i , y i ; s ) is unknown, we consider a model-consistent estimator s of s and use this estimator to

obtain d i = g s ( x i , y i ; s ) , for i s . This leads to the smoothed estimator TySDB , which is obtained by using d i instead
of d i in the design-based estimator TyDB . Note that classical model selection and validation techniques can be used to
determine an appropriate model and to estimate s since we are interested in estimating the relationship between the
design weight variable d and both x and y conditional on the realized sample and only for sample units. We expect that

~
TySDB keeps the good properties of T ySDB in many practical applications provided that the underlying model (1) holds

reasonably well. Indeed, Beaumont (2008) showed that if a linear model holds, we still have that the smoothed estimator

TySDB is unbiased and never less efficient than the Horvitz-Thompson estimator under the model and the sampling
design.
3. WEIGHT SMOOTHING TO HANDLE STRATUM JUMPERS
Let us now consider the weight smoothing approach in the context of stratum jumpers in business surveys. We have
already denoted by Z, the matrix of design information available at the time of the design. Let us denote by Z col , the
matrix of design information at the time of collection, which is assumed to be measured essentially without errors for
sample units. We may hypothesize that, once we know Z col , X and I, the initial design matrix Z brings no extra
3

information about Y. In other words, Y is independent of Z after conditioning on Z col , X and I; i.e.,

F (Y | Z, Z col , X, I ) = F (Y | Z col , X, I ) . This can also be rewritten as F (Z | Y, Z col , X, I ) = F (Z | Z col , X, I ) . The

latter implies that the design weights are independent of Y after conditioning on Z col , X and I and that a suitable model
for the design weights would be
E (d i | Z col , I, X, Y ) = g s (z col , i , x i ; s ) ,
(2)
where z col , i is the vector of design variables at the collection stage for unit i. These design variables are treated here like
additional y-variables. The idea here is to keep from the design weights the useful information contained in z col and x
(since it may have a strong relationship with y) and remove their extra variability. Model (2) can then be used to construct
a smoothed estimator, which should be more efficient than its corresponding design-based estimator.
In many business survey applications, model (2) often reduces to E ( d i | Z col , I, X, Y ) = g s ( z col , i ; s ) . For instance,

X = Z col in the example of Section 4. A simple approach to approximate the unknown function g s (z col , i ; s ) , is then to
discretize z col , i , for i s , into homogeneous categories called collection strata;

this is considered in Section 4.

Assuming that g s ( z col , i ; s ) is constant within each category, we approximate the unknown model by a simple
analysis-of-variance model. The smoothed weight d i is simply obtained as the average of the design weights within the
collection stratum containing unit i. The second-to-last column of Table 2 gives the smoothed weight when the above
methodology is used in the example provided in Table 1.
Table 2. Smoothed weights in the example given in Table 1
Collection
Design
Number of
Design
Smoothed
stratum
stratum
units
weight
weight
A
A
9
1
4
A
B
1
31
4
B
B
40
31
31
50
1280
1280
Sum over the sample units

Smoothed weight (with

constraint)

1 1.0215 = 1.02
4 1.0215 = 4.09
31 1.0215 = 31.67
1253 1.0215 = 1280

For the stratum jumper, the smoothed weight is close to eight times smaller than the design weight. To compensate for
this weight reduction, the smoothed weight of other units in collection stratum A became four times larger than the design
weight. Since units with a small design weight may be associated to large y-values, it is perhaps preferable not to modify
too much the weight of these units as they may become quite influential. One option, tested in Section 4, is to use the
constraint that the smallest design weights are kept unchanged so that the 9 units with a design weight of 1 in Table 2
would also be given a smoothed weight of 1. It may then be necessary to adjust all the resulting smoothed weights by a
constant factor so that the overall sum of the final smoothed weights is still equal to the overall sum of the design weights.
This leads to the last column of Table 2, where the constant factor is 1280 1253 = 1.0215 . This strategy is equivalent to
a hybrid approach between winsorization and weight smoothing, where the largest design weights are winsorized within
each analysis of variance cell. Under this scheme, the winsorization cut-off is simple to compute as it is the average of the
design weights within each collection stratum. Perhaps more sophisticated methods of finding the winsorization cut-off
could yield better results. This has yet to be investigated.
A design-based MSE estimator is

mse(TySDB ) = v(TySDB ) + max 0 , TySDB TyDB

v(TySDB TyDB ) .

(3)

Since TySDB may have a complicated form, the bootstrap technique (e.g., Rao and Wu, 1988; and Rao, Wu and Yue, 1992)
is a natural candidate for obtaining estimators v (TySDB ) and v(TySDB TyDB ) of the design variances Var(TySDB ) and

Var(TySDB TyDB ) respectively. Note that the second term in the right-hand side of (3) is an estimator of the squared
4

design bias, which is restricted to be greater than or equal to zero. Its form is justified in Gwet and Rivest (1992). One can
also restrict the estimated MSE not to be greater than v(TyDB ) since we expect gains in efficiency if model holds. Such
an MSE estimator performed well in the empirical study of Beaumont (2008).
4. AN ILLUSTRATION USING THE CANADIAN WORKPLACE AND EMPLOYEE SURVEY
To illustrate the benefits of the weight smoothing method when handling stratum jumpers, we have used the 2003 CWES
data. The CWES is a longitudinal survey that started in 1999 and that collects information on employers and their
employees. In this application, we focus on the employer portion of the CWES and consider only a subset of its variables.
They are five widely-used financial variables, which will be denoted by Y1 , Y2 , Y3 , Y4 and Y5 . Every other year, a
sample from the population of births is selected from the Business Register. Therefore, the 2003 sample contains units
selected in 1999, 2001 and 2003. In each of these years, employers are selected by stratified simple random sampling
without replacement and the strata are formed by crossing six regions, fourteen industry groups and three size groups. The
size variable corresponds to the number of employees available on the Business Register. To simplify the example, we
restrict to a single region-industry group leading to a sample size of 112.
In this survey, there is a single auxiliary variable xi = z col ,i , with zcol ,i being the number of employees obtained at the
collection stage for employer i. The design-based estimator TyDB is the ratio estimator TyDB =

wi = d i Tx

is

is

wi yi , where

d i xi and Tx = iU xi . The population total Tx is obtained from a reliable external survey (the Survey

of Employment Payroll and Hours) and will be assumed without error to simplify the example.
Figure 1 shows the relationship between the design weights d i and z col ,i . The solid curve has been obtained using the
procedure TPSPLINE of SAS. It is a nonparametric smoothing spline method based on penalized least squares estimation.
We can see that there is a unit with a relatively large design weight of about 35 and a large value of z col . Standard
winsorization of the design weights may not reduce at all the weight of this unit, depending on the cut-off point. Weight
smoothing will be more efficient by reducing the weight of this unit so that it has less influence on the estimates. Note
also that the points are aligned along horizontal lines that represent the different strata.
Figure 1. Plot of the design weights versus z col ,i .

80
60
40
20
0

Design Weight

100

120

(The solid curve is a nonparametric regression fit.)

200

400

600

800
Z_col

1000

1200

1400

To smooth the weights we have used a one-way analysis of variance model with five categories (SR-5) obtained by
discretizing z col as proposed in Section 3. There are three categories for z col 200 , as the slope of the smoothed spline
curve is quite abrupt for small z col and two categories for z col > 200 . The analysis of variance residuals are plotted
against z col and Y4 in Figures 2 and 3. The smoothing splines in these figures do not show obvious trends in the
residuals; thus this model for the design weights is satisfactory. Although they are not provided here, plots of the residuals
against the other y-variables were similar. Thus, the assumption F ( Z | Y, Z col , X, I ) = F ( Z | Z col , X, I ) made in Section
3 is reasonable.
Figure 2. Plot of the analysis-of-variance residuals versus z col ,i .

Residual

(The solid curve is a nonparametric regression fit.)

200

400

600

800

1000

1200

Z_col

Figure 3. Plot of the analysis-of-variance residuals versus Y4,i .

Residual

(The solid curve is a nonparametric regression fit.)

0
Y4

1400

In this empirical study, we have also considered a common mean model (SR-1) for the weights. Under SR-1, the
smoothed weight of all the sample units is simply the average weight. The simulations also feature two estimators, WIN10
and WIN100, obtained by winsorizing the largest design weights with cut-off points of 10 and 100 respectively. Versions
of SR-1 and SR-5 that left the design weights less than 2 ( d i < 2 ) unchanged, as described in the example shown in the
last column of Table 2, were also calculated. All the smoothed weights were winsorized to insure that the final adjusted
weight, say d iF , lies in the range 0.1d i d iF 10d i . This step did not change the results in a significant way; it
controlled the bias by preventing large weight adjustments, especially when the constraint on the smallest design weights
was not used.
For each smoothed or winsorized estimator, Ty* DB say, associated to the design-based ratio estimator TyDB , three
quantities have been computed: i) the relative difference as a percentage, RD = 100 (Ty* DB TyDB ) TyDB ; ii) the Wald
statistic, Wald = (Ty* DB TyDB ) 2 vB (Ty* DB TyDB ) , which can be used to give an indication of the design bias of the
smoothed

or

winsorized

RE_Bootstrap = 100 vB (T

DB
y

estimators;

and

iii)

the

bootstrap

relative

efficiency

as

percentage,

) mse B (T

* DB
y

) . The notation v B (.) is used to indicate that the Rao-Wu bootstrap variance
estimator has been chosen. The estimator mse B (TyRC ) is obtained by using the Rao-Wu bootstrap method for the
estimation of both variance terms involved in (3). One thousand bootstrap replicates have been used in this empirical
study. A summary of the results is reported in Table 3.
Table 3: Comparison of smoothed and winsorized estimators using CWES data
RD
Wald RE_Bootstrap
RD
Wald RE_Bootstrap
Variable Method
Without constraint on
With constraint on
the smallest design weights
the smallest design weights
SR-1
8.88
0.59
434.9 11.96
3.68
56.6
SR-5
-3.77
0.10
323.2 -10.31
0.81
238.4
Y1
WIN10
10.08
1.90
143.8 10.08
1.90
143.8
WIN100
2.34
1.99
92.3
2.34
1.99
92.3
SR-1
10.10
0.62
448.9 14.19
4.05
53.6
SR-5
-5.49
0.16
418.8 -12.98
0.98
318.7
Y2
WIN10
11.70
2.14
135.2 11.70
2.14
135.2
2.53
1.78
92.8
2.53
1.78
92.8
WIN100
SR-1
18.45 10.82
11.3 10.12 20.02
26.5
SR-5
2.22
0.32
155.4
-1.66
0.13
136.4
Y3
WIN10
15.20 16.07
15.6 15.20 16.07
15.6
WIN100
1.72
7.03
88.6
1.72
7.03
88.6
SR-1
137.32 73.97
0.5 29.06 26.68
10.8
SR-5
39.66
6.91
6.3 12.36
3.56
40.5
Y4
WIN10
95.05 61.67
1.1 95.05 61.67
1.1
WIN100
7.30 34.28
64.2
7.30 34.28
64.2
SR-1
47.46
8.65
13.7 27.00 15.33
28.4
SR-5
-13.01
1.26
211.0 -12.22
0.92
215.3
Y5
WIN10
39.47 13.17
18.2 39.47 13.17
18.2
WIN100
5.32
8.60
88.9
5.32
8.60
88.9
From Table 3, we can make the following remarks:

For all variables but Y4 , the SR-5 estimator was not significantly biased, according to the Wald statistic, and it
was more efficient than the ratio estimator. Also, the constraint on the smallest design weights led generally to a
small loss of efficiency.

For variable Y4 , the SR-5 estimator was significantly biased although less biased than its competitors according to
the Wald statistic. This resulted in an inefficient estimator. The use of the constraint on the smallest design
weights substantially reduced the bias and improved the efficiency.
For variables Y1 and Y2 , the SR-1 estimator was the most efficient when no constraint on the smallest design
weights were used, although only marginally more efficient than estimator SR-5. However, imposing this
constraint resulted in a significant loss of efficiency and an increase in bias for these two variables. For the other
three variables, the SR-1 estimator had a very large bias which made the estimator inefficient.
Both winsorized estimators did not perform well. The WIN10 estimator was sometimes significantly biased while
the WIN100 never led to gains in efficiency. We tried several other winsorization cut-offs but were not able to
find any satisfactory compromise. Note also that the constraint on the smallest design weights had no effect on the
winsorized estimators.
Overall, the SR-5 estimator is the best. Also, the constraint on the smallest design weights seems to offer
protection against bias at the expense of a slight loss of efficiency when the bias of the smoothed estimator is not
significant.

Figures 2 and 3 suggest that the analysis-of-variance model used in this example is adequate. Using an argument similar
to Beaumont (2008), the resulting smoothed estimator SR-5 should be asymptotically unbiased and more efficient than the
ratio estimator, under the model and the sampling design, provided that this linear model holds. This is in agreement with
results of Table 3, except for variable Y4 . The bias for variable Y4 may thus be explained either by a slight model
misspecification that is difficult to detect by a graphical analysis or by an error of the Wald test since the Wald statistic is
not that extreme. From a single sample, it is difficult to determine the exact cause of this bias. It is worth mentioning again
that the constraint on the smallest design weights seems to bring some robustness against model misspecification and
potential bias. Another alternative could have been to include Y4 in the model, in addition to z col , in order to reduce the
impact of the possible model misspecification.
As a final comment on this example, note that the SR-1 estimator is equivalent to a model-based ratio estimator when no
constraint on the design weights is used. Such a model-based estimator ignores completely the design weights and should
work well if its underlying model explains satisfactorily the relationship between the y-variables and x. Apparently, this
might have been the case for variables Y1 and Y2 . However, results in Table 3 also indicate that it may be risky in general
to blindly use this estimator unless one is confident that its underlying model holds reasonably well.
5. SUMMARY AND DISCUSSION
We have adapted a weight smoothing method to deal with stratum jumpers in stratified business surveys. The method is
simple to implement as the smoothed weights are simply obtained by taking the average of the design weights within
appropriate collection strata. It has also been shown to be quite promising in our empirical investigation. If weight
smoothing does not yield a sufficient increase in efficiency then nothing precludes, in principle, to combine weight
smoothing with an outlier-robust method, such as winsorization or M-estimation. This remains to be investigated.
It is important to point out that finding an appropriate model is a key aspect of our method. To obtain some robustness
against model failures, we partitioned the sample into collection strata in a more or less ad hoc way. Research into the
issue of determining adequate collection strata boundaries could be useful. An alternative to determine collection strata,
which remains to be investigated, would be to use nonparametric methods of estimating the smoothed weights.
ACKNOWLEDGEMENTS
We wish to thank Harold Mantel and Stphane Tremblay from Statistics Canada for their useful comments and
suggestions. We also wish to thank Cynthia Bocci from Statistics Canada for helping us out with the figures.
REFERENCES
Beaumont, J.-F. (2008). A new approach to weighting and inference in sample surveys. Manuscript under review in
Biometrika.
8

Deville, J.-C., and Srndal, C.-E. (1992). Calibration estimators in survey sampling. Journal of the American Statistical
Association, 87, 376-382.
Gwet, J.-P., and Rivest, L.-P. (1992). Outlier resistant alternatives to the ratio estimator. Journal of the American
Statistical Association, 87, 1174-1182.
Liu, B., Ferraro, D., Wilson, E., and Brick, J.M. (2004). Trimming extreme weights in household surveys. Proceedings
of the Section on Survey Research Methods, American Statistical Association, Alexandria, Virginia.
Pfeffermann, D., and Sverchkov, M. (1999). Parametric and semi-parametric estimation of regression models fitted to
survey data. Sankhya, Series B, 61, 166-186.
Potter, F. (1990). A study of procedures to identify and trim extreme sampling weights. Proceedings of the Section on
Survey Research Methods, American Statistical Association, Alexandria, Virginia, 225-230.
Rao, J.N.K., and Wu, C.F.J. (1988). Resampling inference with complex survey data. Journal of the American
Statistical Association, 83, 231-241.
Rao, J.N.K., Wu, C.F.J., and Yue, K. (1992). Some recent work on resampling methods for complex surveys. Survey
Methodology, 18, 209-217.
Rivest, L.-P. (1999). Stratum jumpers: Can we avoid them?. Proceedings of the Survey Research Methods Section,
American Statistical Association, 64-72.